Few-shot segmentation by enhanced ensemble of base and meta predictions

Date

2023-4-28

Author

Kayabaşı, Alper

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

234
views

160
downloads

Cite This

Supervised learning approaches assume that there is a large amount of data available with labels. Since data annotation is a costly and time-consuming task, this assumption loses its validity when there are financial or time constraints. In addition, only a small number of samples may be available for specific classes, such as images of endangered animals. To address these issues, the concept of few-shot learning has been developed to recognize patterns from novel tasks with limited supervision. As a sub-task of few-shot learning, few-shot segmentation aims to create a generalizing model that can segment query images from unseen classes during training, using a few support images whose class matches that of the query image. Previous research has identified two specific problems in this domain: spatial inconsistency and a bias toward seen classes. To address the issue of spatial inconsistency, the proposed method in this thesis compares the support feature map to the query feature map at multiple scales, making it scale-agnostic. To address the bias towards seen classes, a supervised model called the base learner is trained on available classes to identify pixels belonging to seen classes accurately. The meta learner then uses an ensemble learning model to coordinate with the base learner and discard areas belonging to seen classes. Our method in this thesis is the first to address these two crucial problems simultaneously and achieves state-of-the-art performance on both PASCAL-5i and COCO-20i datasets.

Subject Keywords

Few-shot segmentation, Ensemble learning

URI

https://hdl.handle.net/11511/103126

Collections

Graduate School of Natural and Applied Sciences, Thesis